Research And Implementation Of Action Recognition Method Based On Deep Learning

In recent years,with the development of deep learning and computer vision technology,automatic detection and recognition of human actions in surveillance video have become a hot topic in research and application.Therefore,this thesis studies the deep learning-based action recognition method to detect,locate and identify human actions in the surveillance video scene of public places,such as nursing homes and campuses,and implements a prototype system.The main contributions of this thesis are as follows:(1)For the problem of action localization in long sequence videos,we conduct research and analysis on different action localization methods to find more suitable methods for our system.We primarily introduce two types of existing methods:single-stage and two-stage localization methods.Additionally,we perform experimental analysis on these methods in order to select the more suitable method based on the characteristics.(2)Aiming at action recognition in the video,we propose an actionrecognition method based on human skeletal pose and attention mechanism.This method first extracts the heat map of human skeleton points for fine-grained body action detection,then uses the two-stream 3D convolutional neural network to extract the images and skeleton features.It extracts the spatiotemporal features of video actions,and further optimizes the accuracy of action recognition by an attention mechanism.In order to verify the proposed action recognition algorithm,we design and collect a basic human action dataset.(3)We design and implement a video motion surveillance system based on deep learning.The prototype system has the advantages of high accuracy,convenient use and practical function for action location and recognition.On the benchmark dataset and the basic human action dataset collected,by comparing with other typical methods,the experiments show the effectiveness of our method in surveillance video scenarips.